Native backend: keep the HF exporters working with transformers 4

Summary:
#24422 moved the native HF exporters to the transformers 5 cache API, but some builds still use transformers 4.57. There `StaticLayer.lazy_initialization` takes only the keys, so `PositionedStaticLayer.replacing` raised a `TypeError` and every hf_llm, Whisper and cross-engine export failed. Pass the arguments the installed version takes. `--gguf` now imports `init_empty_weights` from `accelerate`, so add it to the Buck deps. transformers 4 builds Gemma 3's sliding-window layers as `StaticSlidingWindowLayer`, whose KV dtype can't change, so the Gemma 3 fp16 case of `test_cache_writes_index_the_cache_position_input` is skipped in fbcode.

Authored with Metamate (agentcloud).

Reviewed By: shoumikhin

Differential Revision: D130697176

Author: digantdesai

backends/native/examples/hf_llm/BUCK

@@ -28,6 +28,7 @@ fbcode_target(
         "//executorch/exir:lib",
         "//executorch/exir/native:native",
         "//executorch/extension/llm/export:load",
+        "fbsource//third-party/pypi/accelerate:accelerate",
         "fbsource//third-party/pypi/safetensors:safetensors",
         "fbsource//third-party/pypi/transformers:transformers",
         "//pytorch/ao:torchao",
@@ -48,6 +49,7 @@ fbcode_target(
     deps = [
         ":export_hf_llm_lib",
         "//caffe2:torch",
+        "//executorch/export:export_utils",
         "fbsource//third-party/pypi/gguf:gguf",
         "fbsource//third-party/pypi/transformers:transformers",
     ],

backends/native/examples/hf_llm/hf_utils.py

@@ -7,6 +7,7 @@
 """Report native lowering for exported Hugging Face models."""
 
 import collections
+import inspect
 import operator
 
 import torch
@@ -35,9 +36,13 @@ class PositionedStaticLayer(StaticLayer):
     @classmethod
     def replacing(cls, old: StaticLayer, dtype: torch.dtype) -> "PositionedStaticLayer":
         layer = cls(old.max_cache_len)
-        layer.lazy_initialization(
-            old.keys[:, :, :1].to(dtype), old.values[:, :, :1].to(dtype)
-        )
+        keys = old.keys[:, :, :1].to(dtype)
+        values = old.values[:, :, :1].to(dtype)
+        # transformers 4 sizes both buffers from the keys alone.
+        if len(inspect.signature(layer.lazy_initialization).parameters) == 1:
+            layer.lazy_initialization(keys)
+        else:
+            layer.lazy_initialization(keys, values)
         return layer
 
     def update(self, key_states, value_states, *args, **kwargs):

backends/native/examples/hf_llm/test_export_hf_llm.py

@@ -27,6 +27,7 @@
     LastTokenLM,
     load_gguf,
 )
+from executorch.export.utils import is_fbcode
 from transformers import (
     AutoModelForCausalLM,
     Gemma2Config,
@@ -320,6 +321,8 @@ def test_cache_writes_index_the_cache_position_input(self) -> None:
         seq = torch.export.Dim("seq", min=1, max=31)
         for arch, kv_dtype in (("llama", torch.float32), ("gemma3", torch.float16)):
             with self.subTest(arch=arch, kv_dtype=kv_dtype):
+                if arch == "gemma3" and is_fbcode():
+                    self.skipTest("skipping for fbcode")
                 torch.manual_seed(7)
                 model = _tiny(arch)
                 fp32 = LastTokenLM(model, 32).eval()